Origin: deferred bucket (d) from the Tier-3 cr-loop review of PR #167 (merged as 55744ca). Flagged independently by four reviewers. Separate subject from that PR, which only swapped this file's fetch for apiFetch.
apps/web/src/hooks/use-qa-chat.ts consumes the /api/qa SSE stream. Four defects, in severity order.
1. No cross-chunk buffer — frames are silently dropped
:99 does chunk.split('\n') per read, with no carry buffer for a partial frame. A ReadableStream read boundary can fall anywhere, so any SSE event split across two reads is parsed as two invalid halves. Both halves land in the empty catch at :127 and vanish.
The event most likely to be split is the largest one — the metadata frame carrying sources and confidence. So the visible symptom is an answer that renders fine but loses its citations, intermittently, with nothing logged. That is the worst possible shape for a bug: it looks like the model just didn't cite anything.
Fix: keep a buffer across reads, split on \n\n, and retain the trailing partial for the next iteration.
2. [DONE] never matches under CRLF
:100-109 compares against [DONE] exactly. With CRLF line endings the token arrives as [DONE]\r, falls through to JSON.parse, and throws — again into the empty catch. Normalise line endings before comparing.
3. The empty catch is what makes 1 and 2 invisible
:127 swallows every parse failure with no logging. Both defects above would be obvious in a single session with any logging at all. Related: #197 — nothing in apps/web uses the structured logger that exists.
4. Concurrent sends corrupt shared state
:68 assigns a new AbortController over the in-flight one without aborting it, so a second send leaves the first stream running and both write the same state. There is also no abort on unmount. Separately, :61 reads conversationHistory from a stale closure, so rapid sends replay incomplete history to the model — and :152 strands the user's message on abort while the error placeholder gets replayed as assistant history.
Test gaps
No coverage for a frame split across chunk boundaries, no CRLF case, no concurrent-send case. A regression test for #1 needs a mocked stream that deliberately splits a frame mid-JSON — worth writing first, since it fails today.
Origin: deferred bucket (d) from the Tier-3 cr-loop review of PR #167 (merged as
55744ca). Flagged independently by four reviewers. Separate subject from that PR, which only swapped this file'sfetchforapiFetch.apps/web/src/hooks/use-qa-chat.tsconsumes the/api/qaSSE stream. Four defects, in severity order.1. No cross-chunk buffer — frames are silently dropped
:99doeschunk.split('\n')per read, with no carry buffer for a partial frame. AReadableStreamread boundary can fall anywhere, so any SSE event split across two reads is parsed as two invalid halves. Both halves land in the emptycatchat:127and vanish.The event most likely to be split is the largest one — the
metadataframe carrying sources and confidence. So the visible symptom is an answer that renders fine but loses its citations, intermittently, with nothing logged. That is the worst possible shape for a bug: it looks like the model just didn't cite anything.Fix: keep a
bufferacross reads, split on\n\n, and retain the trailing partial for the next iteration.2.
[DONE]never matches under CRLF:100-109compares against[DONE]exactly. With CRLF line endings the token arrives as[DONE]\r, falls through toJSON.parse, and throws — again into the empty catch. Normalise line endings before comparing.3. The empty catch is what makes 1 and 2 invisible
:127swallows every parse failure with no logging. Both defects above would be obvious in a single session with any logging at all. Related: #197 — nothing inapps/webuses the structured logger that exists.4. Concurrent sends corrupt shared state
:68assigns a newAbortControllerover the in-flight one without aborting it, so a second send leaves the first stream running and both write the same state. There is also no abort on unmount. Separately,:61readsconversationHistoryfrom a stale closure, so rapid sends replay incomplete history to the model — and:152strands the user's message on abort while the error placeholder gets replayed as assistant history.Test gaps
No coverage for a frame split across chunk boundaries, no CRLF case, no concurrent-send case. A regression test for #1 needs a mocked stream that deliberately splits a frame mid-JSON — worth writing first, since it fails today.